Stanford University AI Guidelines for Marketing and Communications: Responsible Use and Best Practices

Within the context of these guidelines, 'Artificial Intelligence' or 'AI' encompasses a broad spectrum of software systems. This includes tools that generate content such as text, images, video, audio, and code, as well as those that analyze data, automate tasks, or provide recommendations. These capabilities are typically powered by machine learning, large language models (LLMs), and related advanced techniques. Prominent examples include widely used platforms like ChatGPT, Claude, Gemini, and Copilot. Furthermore, the definition extends to AI features embedded within existing software, such as advanced grammar checkers, presentation assistants, and analytics platforms. It also covers custom applications, agents, and automated workflows developed using AI models or APIs. As a general rule, any tool that generates, predicts, or recommends outputs through systems trained on data, such as neural networks or LLMs, falls under the purview of these guidelines for marketing and communications activities. When in doubt about a tool's classification, consulting with a manager, the IT organization, or University Communications is strongly advised. It is important to note that these guidelines specifically address the application of AI within marketing and communications functions and do not provide training on how to use individual AI tools. Resources for such training are available through University IT and University Communications' training programs. The absence of a specific use case in this document should not be interpreted as a prohibition; Stanford favors flexibility, iteration, responsible experimentation, and continuous growth in AI adoption.

The way prompts and other inputs are handled, used, and stored by AI tools is a critical consideration. Ensuring that these interactions meet the appropriate data classification and security requirements is paramount for protecting Stanford's information assets. The university's risk classification framework categorizes data into high-, moderate-, and low-risk tiers, each with specific requirements for AI use. **High-Risk Data:** This category includes sensitive information such as protected health information (PHI), student education records, donor information, and employee personnel records. High-risk data may only be used in environments explicitly approved for such data. This means that unless a specific environment has been sanctioned for high-risk data, it cannot be used in prompts, as attachments, via APIs, or in any other interaction with an AI tool. There are no exceptions to this rule; high-risk data usage is strictly confined to approved environments designed for that specific data category. If there is any uncertainty about whether data qualifies as high-risk, users should consult the risk classification guide or contact their IT organization. **Moderate-Risk Data:** For moderate-risk data, its use with AI tools is permitted only when the tool provides appropriate protections. This typically involves utilizing university-provisioned tools or those that have been reviewed and approved by the IT organization for handling moderate-risk data. Users are expected to exercise good judgment and meticulously document the safeguards implemented. **Low-Risk Data:** Low-risk data, such as publicly available information, can be used with AI tools in accordance with these guidelines. However, even with low-risk data, it is essential to remain mindful of the tool's data retention policies, including those concerning user prompts, and its training practices. **Protecting Stanford Content from Model Training:** A significant concern is the potential for some AI tools to retain user inputs and use them for model training or improvement. Before submitting any Stanford content—including unpublished articles, brand assets, strategic documents, internal research, or institutional data—to an AI tool, users must confirm that the tool's policies explicitly prevent the use of such inputs for training. University-provisioned tools, such as the AI Playground, are specifically configured to prevent this. For other tools, a thorough review of the provider's data use and retention policies is imperative.

Choosing the right AI tool is crucial for effectively serving work needs while simultaneously protecting the university's data and interests. Stanford provides guidance to assist in making informed decisions. **Prioritize University-Provisioned Tools:** The UIT AI Playground is the recommended starting point for most common AI tasks, including drafting, research, brainstorming, editing, and summarization. This platform offers access to a range of large language models within an environment specifically configured to safeguard Stanford data. Files uploaded to the Playground are not shared externally or used for model training, making it the lowest-risk option. **Using Other Tools:** When the capabilities of the AI Playground are insufficient, other AI tools may be considered. Stanford maintains a GenAI tool matrix that compares approved options and their suitability for various data types. If considering an AI tool not on this matrix, it is imperative to verify and understand the following critical criteria before use: * **Data Retention:** Ascertain whether the provider retains user inputs, for how long, and if there is an option to opt out of data retention. * **Training on Inputs:** Determine if the provider uses user inputs to train or improve its models. Investigate if an enterprise or professional tier is available that prevents such training. * **Terms of Use:** Review the terms to ensure they do not grant the provider rights to use Stanford's name or content in their marketing (prohibited under Administrative Guide 1.5.4). Also, check if the terms grant broad licenses over user input, content, or outputs. * **Security Posture:** Verify if the provider offers encryption in transit and at rest, and if they maintain relevant security certifications such as SOC 2, ISO 27001, or equivalent. * **Data Classification:** Confirm that the tool is appropriate for the sensitivity of the data intended for use, referencing the 'Data classification and security requirements' section. If these criteria cannot be verified, the tool should not be used with Stanford data, regardless of its risk level. Ideally, the evaluation of such tools should be conducted in partnership with the local IT team or University IT.

AI agents and automated workflows represent a more advanced application of AI, differing significantly from standard interactions with language models. Instead of manually typing prompts and reviewing responses, these systems can perform actions autonomously on behalf of users. This includes tasks such as monitoring feeds, generating and scheduling content, sending alerts, updating databases, or executing multi-step processes with minimal or no real-time human intervention. Stanford's Information Security Office has published specific guidance on agentic AI that is applicable to these use cases. The following principles are critical for the responsible use of agents or automation processes: * **Close Collaboration with Your IT Organization:** Any agent or automated workflow operating within a Stanford environment must be developed and maintained in direct collaboration with the local IT team. This applies to agents that monitor data sources (e.g., daily news digests), execute scheduled tasks (like data retrieval), or interact with Stanford systems. * **Human Oversight Scales with Stakes:** The level of human oversight required should be directly proportional to the potential consequences of the agent's actions. An agent that flags potential stories for editorial review, for instance, carries a lower risk than one that publishes content directly to a Stanford website. For any automated workflow that performs an action visible to an external audience—such as publishing, sending, or posting—a human must review and approve each action before it is executed. * **Logging and Auditability:** Automated workflows must maintain comprehensive logs that are sufficient to understand precisely what actions were taken, when they occurred, and on what basis (i.e., the inputs used). This is crucial for both troubleshooting operational issues and ensuring accountability. * **Scope Limitation:** Agents should be designed with the narrowest possible scope of authority necessary to accomplish their intended task. For example, an agent built to monitor academic publications should not possess the capability to post content or modify data in other unrelated systems. This principle of least privilege helps mitigate potential risks and unintended consequences.

Stanford University's commitment to innovation is matched by its dedication to responsible implementation. These AI guidelines for marketing and communications provide a robust framework for harnessing the power of artificial intelligence while upholding the university's core values, legal obligations, and ethical standards. By emphasizing human oversight, alignment with institutional goals, and transparency, Stanford aims to foster an environment where AI serves as a valuable tool to augment human capabilities, drive efficiency, and enhance communication efforts. The guidelines encourage thoughtful experimentation, continuous learning, and a proactive approach to data security and intellectual property. As AI technology continues to evolve, Stanford remains committed to adapting its practices, ensuring that its use of AI is always in service of its mission and its community. The principles outlined here are not static but represent an ongoing commitment to navigating the complexities of AI with integrity and foresight.